Why are agents paying twice as often for smaller amounts?
In the four and a half days this issue covers, the sellers we track were paid about
twice as often as the week before — and the average payment shrank from $3.50 to
$2.26. More payments, smaller payments: that is the week's shape, and it showed up
in the categories where agents make lots of tiny purchases. Also this week: hotel
booking went live inside Google's AI Mode, Amazon's shopping agent learned standing
watch-and-alert tasks, and a new payment primitive for agents launched on Solana.
Updated weekly · data as of 2026-10-13 · window 2026-08-31–2026-09-04 · method score-v1.0.0
· tracked-catalog data · methodology
What agents did this week
Every figure in this section is recomputable via the published score-v1.0.0 method
and recorded in this issue's claims ledger.
Scope. These are tracked-catalog numbers —
USDC received on
Base by the
1,1076 seller
addresses in our x402 Bazaar-derived catalog
(8796 tier-A) —
never "the x402 market" or "the network." Wallets can receive USDC for reasons
unrelated to x402 sales; the method removes only known exchange, fan-out, and self
flows (which removed just
~$242.651 this
window).
What "verified" means here. "verified" here is the payment-level
predicate of score-v1.0.0 (a payment of at least $0.01, sender not flagged
exchange/fan-out, not a self-transfer). It is not the seller-level
"verified tier." The two labels must not be conflated in copy.
Every number in this section is verified on-chain — you can recompute it.
Dust, shown beside verified and never netted out:
across the whole window the catalog saw
122,7204 transfers
gross, of which
65,9074 were dust
moving $218.534
together; gross volume including dust and excluded senders was
$127,504.094 (a
context number, never labeled verified). Verified volume for the whole window, tail
included: $127,261.441
across 56,3441
payments.
The pace doubled in counts, rose a third in dollars. On the three
fully published days of this window, verified volume ran
$34,129.82/day1
against $25,808.00/day2
the prior full week (08-24 → 08-30:
$180,656.022 across
51,6732 payments) —
a +32.2%2 volume
pace. Payment frequency roughly doubled:
14,8681 verified
payments/day vs 7,3822 /day.
The mean verified payment fell from
$3.502 to
$2.261 . More
payments, smaller payments.
The tail is a floor. 2026-09-03 recomputes to
$16,607.743
(10,5233 payments)
and the partial 2026-09-04 to
$8,264.233
(1,2173 payments)
by the same method, but these days are not yet coverage-published and some tier-A
cursors (oldest 2026-09-03T14:00Z) have not passed the 09-03 day-close; tier-B sellers
sync on a 7-day cadence, tier-C on 30 days. These figures WILL revise upward.
market_daily is append-only, so any revision is visible.
Who was paid.4551 distinct
sellers were paid this window and
1,1561 distinct
senders paid them (vs
5962 and
1,9362 across the
prior full week — unequal window lengths, so no raw comparison is printed
without the rate framing).
381 sellers
received their first-ever observed verified payment this window
(~8.3/day1 vs
11.0/day2 prior
week). "First-ever" means first in our fully backfilled history for cataloged
addresses; an address earning before we cataloged it is dated by chain history, and
payments to addresses never cataloged are invisible by design.
Where the money went. The top three categories by verified volume
were the same both weeks and dominant: payments-infrastructure
($108,645.335 ;
3,2995 payments;
845 sellers),
search-retrieval
($104,540.545 ;
41,0745 ;
1685 ), and
storage-state
($86,645.865 ;
4,5425 ;
455 ). data-access
overtook inference for fourth place with absolute growth —
$19,076.265 in
~4.5 days against
$14,588.195 across
the prior full seven. The falling mean ticket shows up as a high-frequency,
small-ticket surge in payment counts: media-generation
36,7735 payments vs
22,6785 prior week,
communications
36,4025 vs
19,6605 ,
scraping-extraction
38,2475 vs
23,6195 — each in
fewer days. Category disclosure: a seller listed in several categories
counts in each of them (the frozen v_category_market convention), so category sums
exceed the deduped window total
(~$359.8k5 of
category volume vs
$127.3k5 deduped).
This is disclosed, not a bug.
How people used agents this week
Hotel booking went live inside Google Search's AI Mode
move
People in the US can now book a hotel end-to-end inside Google Search's AI Mode:
describe the stay in plain English, compare options, pick a room from roughly 10
launch partners (including Choice Hotels, Hilton and Wyndham), and complete the
reservation with Google Pay without leaving the conversation. New delegation:
booking a hotel inside a general search agent, not just researching one.
who Google
date announced 2026-08-27; confirmed
operating in trade coverage 2026-09-01 (Hotel Dive) — announcement date is
pre-window, disclosed
changes our ledgers?Yes — our /economy ledger is updated. The
Google UCP checkout
(AI Mode and Gemini) entry now carries live US hotel booking with Google Pay,
both dates printed. The entry's status stays limited, and Google's own "has started
rolling out" wording is carried rather than a flat "live."
Amazon launched "Update Me When" in Alexa for Shopping for US customers: tell the
agent to watch for things you care about — a brand's product-line launch, an
author's next book, an artist's tour, a new TV season — and it alerts you when it
happens. It joins Scheduled Actions, which already adds recurring items to your
cart for review before checkout. The person still completes the purchase.
changes our ledgers?No status flip. Matches the standing event trigger "Amazon changes
its shopping surface (expansion)" on
/economy/buying-from-businesses;
the existing "Amazon Buy for Me is limited" entry stands.
How many people let an AI finish the purchase
number
11% of surveyed US consumers say they have used AI to complete a transaction —
against 53% who used AI for research in the past month and 33% for product
discovery. 76% are uncomfortable letting AI complete purchases on their behalf;
63% would let it auto-insert discount codes at checkout. Parents are nearly three
times as likely to have completed a transaction with AI (19% vs 7% of adults
without children at home).
who NMI (an embedded-payments provider),
"Embedded Payments: Agentic Commerce Reality Check," reported by WWD via Yahoo Tech
counts vs dollars Counts self-reported
shares of a 1,000-consumer US poll — who says they did or would allow each
task. Survey behavior and comfort levels, not transaction counts or dollars.
changes our ledgers? No.
Shoppers arriving via AI tools spend more per visit
number
AI-associated referral traffic to US ecommerce sites rose 62% year over year in
July, and shoppers arriving via AI tools generated 53% more revenue per visit than
other shoppers. AI referrals still have not overtaken traditional search referrals
in volume.
who Adobe Analytics, reported by Digital
Commerce 360
date 2026-09-03 (July data, published this
week — disclosed)
counts vs dollars Counts referral visits
to US retail sites tracked by Adobe Analytics, and revenue per visit — growth rates
and per-visit comparisons, not absolute dollars.
changes our ledgers? No.
Which assistant people shop through is shifting
number
ChatGPT was the top AI referral source for 722 of North America's top 1,000 online
retailers in Q2 2026, down from 844 in Q1. Google's Gemini doubled from top source
for 16 retailers to 32, Perplexity gained 12 retailers, and Claude went from 1
retailer to 15.
who Digital Commerce 360 and ReFiBuy
(AI Commerce Rankings; Scot Wingo, ReFiBuy CEO, quoted)
date 2026-09-03 (Q2 data, published this
week — disclosed)
counts vs dollars Counts
retailers (from DC360's Top 1000 by web sales) ranked by which assistant
sends each the most referral traffic, Q2 vs Q1 2026 — a count of retailers, not
visits, transactions, or dollars.
changes our ledgers? No.
The big moves
Anthropic released an open-source commerce-agent blueprint
Anthropic released Claude Commerce Agents, a free open-source blueprint
(Apache-2.0, on GitHub) for retail, travel, telecom, and ticketing companies: a
shopping agent that searches a catalog, compares options, answers order questions,
and fills a customer's cart inside the business's own site or app, plus a merchant
agent for inventory, prices, and campaigns where every change waits for a person's
approval. The blueprint completes no payments itself — checkout stays with the
business's own systems.
who Anthropic; press coverage names
Shopify, Visa, Mastercard, and Accenture as early partners/users
date 2026-09-02 (per Reuters; some
secondary coverage says 09-03 — Reuters date used)
reported figures Anthropic's pilot
figures — cart size up about 30–35% for one partner, purchase completion about 60%
likelier — are vendor-reported with no published methodology, not independently
verified. Shopify's claim that its reference implementation connecting the shopping
agent to Shopify stores was already live rests on a LinkedIn post by Shopify's
president — a reported claim, not independently verified.
changes our ledgers?Yes — our /economy ledger is updated.Claude Commerce
Agents is now an entry there (live — open-source blueprint), worded for what it
is: build tooling, not a payment surface; it completes no checkout of its own.
Payment Channels went live on Solana
Solana launched Payment Channels, a payment primitive built for AI agents: a
spending limit is authorized once on-chain, the agent runs many tiny payments
against it off-chain (Solana's own analogy: a bar tab, or a prepaid meter), and the
total settles in a single on-chain transaction. Alibaba Cloud is the launch
partner — Solana says Alibaba Cloud API endpoints are live with Payment Channels at
launch, meaning an agent can meter-pay for those cloud services today.
who Solana (Solana Foundation), with
Alibaba Cloud as launch partner
counts vs dollars Solana's "1 million
payments per second" is the vendor's own testing figure, and it counts payments —
not dollars moved.
changes our ledgers?Yes — our /economy ledger is updated.Solana Payment
Channels now sits in the payment rails and protocols section (live, 2026-09-03).
MoonPay's PayBox went live inside X's Grok
MoonPay's PayBox — a payment vault for AI assistants — went live inside X's Grok,
after earlier launches in Claude and ChatGPT. The assistant prepares a transaction
(buy or swap crypto, bridge funds, pay over the x402 protocol, or use integrations
for travel bookings, restaurant reservations, and shopping) and the user approves
it with a passkey before any money moves — or grants an "Autonomous" mode where the
AI acts within limits the user sets. Wallet keys are split so neither MoonPay nor
the AI can sign alone, and access is revocable.
caveat From MoonPay's own announcement:
the vault is operated by Launchpad Limited, a B.V.I. company MoonPay itself
describes as unregulated.
changes our ledgers?Yes — our /economy ledger is updated.MoonPay PayBox is now an entry on the
equipping-your-agent page (live — in Grok, Claude, and ChatGPT, per MoonPay), with
the Launchpad Limited caveat printed in the entry.
Where are the current numbers?
This issue is a dated document — its figures are frozen to 2026-09-04 and
recorded in its claims ledger. The live pages keep moving:
the market dashboard for daily volume,
the rankings for who earns verified revenue, and
the directory for the tracked catalog.
This window's verified figures: every USDC
transfer on Base to a tracked catalog address
between 2026-08-31 00:00 UTC and the observation cutoff, passed through the
score-v1.0.0 verified predicate (at least $0.01, sender not flagged
exchange/fan-out, not a self-transfer). Recomputed from the warehouse record
observed through 2026-09-04T13:13:19Z, block
50,869,126.
Methodology: verified revenue
Prior-week comparison figures: the full week 2026-08-24 → 2026-08-30 at
coverage_ratio 1.0 — $180,656.02 across 51,673 verified payments. The windows are
unequal lengths, so every comparison prints as a daily rate, never a raw total
against a raw total. Same method and observation as note 1.
Tail-day figures: 2026-09-03 and the partial 2026-09-04 recompute by the
same method but are not yet coverage-published — they are floors that will revise
upward as slower-syncing sellers catch up, visibly, in the append-only market_daily
record at /api/v1/market/daily.
Payments below the published dust
floor, counted and summed for the window; displayed beside verified, never netted
out. Gross volume includes dust and excluded senders and is a context number, never
labeled verified. Methodology: dust floor
Category figures follow the frozen v_category_market convention: a
seller listed in several categories counts in each of them, so category sums exceed
the deduped window total. Disclosed wherever category figures print.
Catalog and coverage at observation: 1,107 tracked seller addresses,
879 of them tier-A. Chain-sync last success 2026-09-04T13:52:51Z; all 879 tier-A
sellers sync_status=complete; 0 tier-B/C sellers unsynced since window start. Prior
week fully covered at coverage_ratio 1.0 and reconciles to the micro
(180,656,016,789) between the transfers recompute and the published v_market_daily
weekly sum. Buyer flags are applied at their current version to both windows,
identically to the pipeline itself; a future flags bump can revise past weeks,
visibly.
The figures in "What agents did this week" are chain-derived —
recomputable from public USDC transfers on
Base — and frozen to this issue's date. Reported items in the other sections carry
their own sources inline and are labeled for what they are. Machine-readable source:
/api/v1/market/daily